Digital Twin Neural Marker Discovery for Delineating Mixed Dementia with Cross‐site Federated Learning

Abstract

BACKGROUND: Mixed dementia is characterized by significant clinical and pathological heterogeneity posing challenges for accurate diagnosis and treatment planning. Digital twin technology offers a transformative approach by learning data characteristics to simulate predictions of an unseen data sample. This study aims to construct a digital twin neural marker with multimodal neuroimage data to enable predictive tasks on samples with mixed dementia. We leverage cross-site neuroimage datasets from the Dementias Platform UK (DPUK) and a Korean dementia cohort to construct the digital twin neural marker and show that the marker can be applied to delineate etiologies of mixed dementia. METHODS: A digital twin neural marker of mixed dementia was constructed by training a Vision Transformer (ViT) model with a generative pre-training approach. Specifically, Masked AutoEncoder (MAE) training of the ViT model on T1w and FLAIR neuroimage data from DPUK and a Korean dementia cohort was first constructed to capture the latent characteristics of the neurodegenerative brain etiologies. The model was further fine-tuned on another Korean dementia cohort data with mixed dementia to delineate the etiologies of the neurocognitive disorder given a neuroimage input. Federated learning approaches were employed to train the model without data transfer across sites, maintaining data privacy within the trusted research environment. RESULTS: The fine-tuned digital twin neural marker achieved a reliable performance in delineating the mixed dementia etiology. The latent representation of the neural marker showed a separable pattern between different underlying etiologies when visualized with UMAP. These findings demonstrate the effectiveness of digital twins in leveraging global cross-site datasets to provide actionable clinical insights. CONCLUSIONS: The results highlight the potential of a digital twin neural marker to address the complexities of delineating mixed dementia etiologies. Most importantly, this work represents an international collaboration (Korea-UK) to develop a digital twin neural marker using large datasets while upholding data democracy principles.

Ähnliche Arbeiten